Re: [PATCH 2/2] net: thunderbolt: Update ring indices only after all frames are queued

[Mika Westerberg <mika.westerberg@linux.intel.com>]

Hi,

On Sat, Sep 05, 2026 at 01:57:02PM -0700, Jakub Kicinski wrote:
> On Wed,  2 Sep 2026 10:21:28 +0200 Mika Westerberg wrote:
> > Take advantage of the new API and update the ring producer/consumer
> > indices only after the frames have been queued. This reduces the number
> > of doorbell register writes from one per frame to one per batch.
> > 
> > Measured using iperf3 over a Thunderbolt/USB4 net connection between
> > Arrow Lake and Panther Lake systems (average of three runs):
> > 
> >   before         after
> >   ~14.3 Gbit/s   ~15.3 Gbit/s
> 
> Acked-by: Jakub Kicinski <kuba@kernel.org>

Thanks!

I applied these two to thunderbolt.git/next.

> but Clashiko (sort of) hints at lack of proper handling of transient
> memory allocation failures. Indeed the NAPI path seems to call
> tbnet_alloc_rx_buffers() which flushes the ring completely on
> any sign of trouble. (a) aren't we freeing buffers already handed
> to HW; (b) if we flush Rx free list and there's no Tx won't the driver
> wedge Rx and never schedule NAPI?
> 
> I'd suggest fixing that first before making any optimizations :S

I would have fixed these prior had I known they existed ;-) I will look
into these.